Intel GPU support

[bjia56]

This is a possible implementation for #644. It uses stripped down code from igt-gpu-tools to handle requests to the Linux perf framework and access metrics from the Intel PMU in the same way that intel_gpu_top works. Tested on an Intel Gemini Lake processor's iGPU.

Much of the line count comes from sources pulled (and modified) from igt-gpu-tools, including a database of Intel chipsets and their corresponding pretty-print names. Extra dependencies that would have been pulled in by igt-gpu-tools were removed to keep btop standalone.

[ingenarel]

oh my god i was searching how to have a gpu graph. currently using igpu-tools. would love to see this merged!

[bjia56]

If folks are interested in testing out this change, I'm building downloadable btop zips (with the Intel support) here: https://github.com/bjia56/btop-builder

My use case is to run btop as part of a Docker or LXC container as part of the Scrypted home video platform. Using the Intel PMU requires privileged/root permissions, so the above builds also contain a workaround for pulling GPU metrics from an http server running on the host.

[aristocratos]

@bjia56
Nice work!
Let me now when you think it's ready to go from a draft to a "sharp" PR and I'll start going trough it.

I'm assuming there aren't any conflicting licensing from the code you pulled from igt-gpu-tools?

[bjia56]

I can clean things up a bit and fix the Makefile (have only been using the CMake builds myself).

Is the http-based metrics endpoint (as a fallback data source) of interest for upstreaming, or would you like to keep only the Intel PMU collection method? The tradeoff is the http endpoint is non-standard, while the PMU requires sudo (or at least CAP_PERFMON permissions). For something like Scrypted, which can be internet-facing, we'd like to minimize the permissions required to run containers.

Regarding licensing, to my understanding igt-gpu-tools is MIT-licensed: https://gitlab.freedesktop.org/drm/igt-gpu-tools
The part that I'm unsure of is the DRM headers, as those appear to be pulled from the Linux kernel repo. I am not a lawyer, but since these sources retain their licensing headers, perhaps that's ok?

[aristocratos]

@bjia56

Is the http-based metrics endpoint (as a fallback data source) of interest for upstreaming [...]

I rather avoid pulling in an entire http library for a single use case. As long as we make sure the CAP_PERFMON permission/sudo requirement is well documented. Could maybe add a make setcap to the Makefile that runs setcap on the binary with the correct arguments to make it even easier. Could also be a good idea to log a WARNING when starting up if permissions is missing.

Regarding licensing, to my understanding igt-gpu-tools is MIT-licensed [...]
The part that I'm unsure of is the DRM headers [...]

Looks like there is a MIT notice on the DRM headers also, so if that is the case from the origin it shouldn't be any issue.

It's only a problem if it was GPL, since GPL don't mix with Apache.

[aristocratos]

Nice work @bjia56 :)

Merging in to main.

[bjia56]

Thanks for finishing up this PR!

[RudolphSedlin]

Are we certain that this works for Alchemist GPUs using the i915 KMD? I would assume so given the KMD is the same, and yet my A770 is unrecognized. Running in debug mode suggests no GPU collect occurs at all.

[bjia56]

Does the A770 show up in intel_gpu_top?

[RudolphSedlin]

Yeah, its shown as just "Intel Dg2 (Gen 12)". Hardware device is also as expected, just /dev/dri/card1. This is rather bizarre, given the codebase should be the same between the two, though it should be worth noting I cleanbuilt igt-gpu-tools, perhaps btop is out-of-date somehow?

[bjia56]

The reference igt code that was pulled into btop was commit 0f02dc176959e6296866b1bafd3982e277a5e44b from https://gitlab.freedesktop.org/drm/igt-gpu-tools.git. btop uses the PMU to access metrics, so it requires sudo permissions same as igt. If the device itself is unknown by the built-in Intel chip database, then it should show as a Generic device, if I recall correctly.

[andras-tim]

I compiled btop recently with GPU_SUPPORT=true, and then with sudo I saw my Intel GPU. Thanks!

Edit: I have an old, integrated GPU:

$ intel_gpu_top -L
card0                    Intel Kabylake (Gen9)             pci:vendor=8086,device=591B,card=0
└─renderD128

[RudolphSedlin]

How odd, then the device is not generic. Tried to compile with GPU_SUPPORT=true, no dice. May try with Xe, maybe there is something there. @andras-tim, do you get metrics on your Alchemist card?

Edit: No luck, I assumed perhaps a different KMD would expose the PMU differently if the fault was somehow localized to my machine.

[luyu-wu]

Is there anyway to do this without root? I believe nvtop (and missioncenter) gets basic usage stats without root permissions!
I'll try and find some time to look into it if it's possible (and applicable here).